The short answer
Pick AgentVibes for free voices on any OS; pick Heard for updates that do not depend on the agent remembering to speak.
AgentVibes is a free, Apache-2.0 tool that gives Claude Code a spoken voice. A Claude Code hook adds instructions (on every prompt on macOS and Linux, at session start on Windows) that tell Claude to run AgentVibes’ speech script with a short line when it acknowledges a task and another when it completes it. It can set up Codex and GitHub Copilot the same way through its MCP server, and ships a hook for Hermes. It has hundreds of voices (local Piper, Kokoro, macOS Say, Windows SAPI, or ElevenLabs in the cloud), runs on macOS, Linux and Windows, and can play speech on your laptop when the agent runs on a remote server.
Heard (heard.dev), the voice layer for coding agents, works from outside the agent. It reads Claude Code and Codex hook events and decides what deserves speech, so questions, approvals and failures are spoken even if the agent never calls a tool, and in the app, routine updates from several sessions are batched by project. It runs on macOS 14 or later; the open-source engine is free and the Mac app has paid plans. AgentVibes wins on price, platforms, voice choice and remote servers.
Side by side
What each one does.
AgentVibes facts from its README, docs and release notes (v5.15.2). Checked October 22, 2026.
| Tool | When it speaks | Who writes the words | Voices | Agents | Platforms | Price |
|---|---|---|---|---|---|---|
| AgentVibesnpx agentvibes | When a task starts and when it finishes (Hermes: every response) | The agent, following instructions an AgentVibes hook adds to the session | Piper (900+), Kokoro, macOS Say, Windows SAPI, Soprano; ElevenLabs optional | Claude Code (hooks); Codex and GitHub Copilot through MCP plus instructions; Claude Desktop and Warp through MCP; OpenClaw; Hermes through a hook | macOS (needs bash 5), Linux, Windows, Android via Termux | Free, Apache-2.0 |
| Heardvoice layer | Questions, approvals and failures in every mode; results and routine steps by mode (Focus keeps both quiet); several sessions batched by project in the app | Heard, from agent hook events | Local Kokoro or your own keys (engine); managed voices (app) | Claude Code, Codex CLI and Codex app; session status for Cursor | macOS 14+ | Engine free, Apache 2.0; app Pro $15/mo, Power $30/mo, 14-day Power trial |
What AgentVibes does well.
AgentVibes is about the voice itself. Its interactive installer (npx agentvibes install) wires up the Claude Code hooks and walks you through picking a voice. From there you can stack reverb, echo and chorus, play background music under the narration, apply a personality such as pirate or sarcastic, narrate in another language, and give each agent on a BMAD team its own voice. More than 50 slash commands have plain-English equivalents through its MCP server.
Two features stand out. The Remote Receiver plays speech on your local speakers when the agent runs over SSH on a box with no audio. And sessions stay quiet unless you turn on their project, which also keeps the instructions and their token cost out of projects you have not enabled.
- Free and open source (Apache-2.0).
- macOS, Linux, Windows and Termux.
- Hundreds of voices, effects and music.
- Speech from remote servers on your laptop.
brew install bash # macOS only, once: AgentVibes needs bash 5
npx agentvibes installWhat Heard does differently.
Heard sits outside the agent. It installs hooks for Claude Code and Codex and turns their events into a sentence or two, so the agent spends no tool calls or tokens on talking and nothing changes in your prompts. Routine tool calls get a one-liner or stay quiet, depending on the listening mode. Questions, approvals and failures come through in every mode, and results do too in Co-pilot and Companion; Focus keeps routine results quiet. In the app, with several sessions running, routine updates are batched into short summaries that start with the project name.
The Mac app adds managed voices and spoken recaps on Pro. On Power you hold right ⌘ to answer an agent by voice, ask Parrot about a repo, and follow along on the iPhone companion. The open-source engine is a separate Apache 2.0 build you run yourself with local Kokoro voices.
- No instructions added to the agent.
- Questions and approvals spoken as they happen.
- Summaries across sessions, project name first.
- Voice replies and iPhone on Power.
Pick AgentVibes for voices, platforms and remote boxes.
If you are on Linux or Windows, Heard is not an option and AgentVibes is. It is free, has hundreds of voices plus effects and background music, and its Remote Receiver solves a real problem for anyone who runs agents on a server. If you like the agent describing its own work in its own words at the start and end of each task, that is exactly what it does.
Pick Heard when you need to hear about questions and approvals.
Because AgentVibes’ speech comes from the agent following its instructions, what you hear depends on the agent running the speech command at the right moments. Heard reacts to the agent’s own hook events, so a permission prompt or a failure is spoken when it happens. It reads Codex events directly too, instead of asking Codex to call a speech tool, keeps several sessions apart by project, and on Power lets you answer back by voice.
Running them together.
Both would speak about the same Claude Code session, so you would hear two voices. If you want AgentVibes for a remote server and Heard on your Mac, leave AgentVibes off for the projects you run locally; it stays quiet in any project you have not enabled, unless you have turned it on globally.
Questions
Frequently asked questions
Is AgentVibes free?
Yes. AgentVibes is open source under Apache-2.0 and installs with npx agentvibes install. Its free voice providers need no API key; ElevenLabs is optional and billed by ElevenLabs.
How does AgentVibes decide what to say?
In Claude Code, its hook (UserPromptSubmit on macOS and Linux, SessionStart on Windows) adds a short protocol to the session that tells Claude to run AgentVibes’ play-tts.sh script twice per task: once to acknowledge it and once to report the result, each under 150 characters. Claude writes those lines, and a verbosity setting decides how much reasoning they include. In Codex and GitHub Copilot, an instructions file tells the agent to call its text_to_speech MCP tool at the same two points. Claude Desktop and Warp can call that tool too, but only when you ask them to speak.
Does AgentVibes work on Windows and Linux?
Yes. It runs on macOS, Linux and Windows, and on Android through Termux. On macOS it needs bash 5 (brew install bash). Heard is macOS only, on macOS 14 or later.
Does AgentVibes work with Codex?
Yes. Its setup screen has an OpenAI Codex option that adds AgentVibes’ MCP server to .codex/config.toml and an AGENTS.md telling Codex to call its text_to_speech tool when it starts and finishes a task, the same pattern as in Claude Code. GitHub Copilot gets the same MCP-and-instructions setup. Heard reads Codex CLI and Codex desktop app hook events directly, so it does not depend on Codex following instructions.
What is the main difference between Heard and AgentVibes?
AgentVibes has the agent speak its own short lines at the start and end of each task, with a wide choice of free voices on any OS. Heard (heard.dev), the voice layer for coding agents, speaks for the agent from its hook events on macOS, including questions and approvals, and in the app batches routine updates from several sessions by project.
Ready when your agents are
Hear the work without watching every screen.
Start with the full Heard Power experience for 14 days. No card required. The open source engine remains free.
Download Heard for Mac macOS 14 or later · Claude Code and Codex supported